6.68 · MCP review — Tổng kết pattern

MCP architecture summary

┌───────────────────────────────────────────────┐
│                                               │
│   YOUR APP                                    │
│   ├─ Anthropic client (Claude)                │
│   └─ MCP client(s)                            │
│       │                                       │
│       ├─ Connect via stdio/HTTP/WS            │
│       ▼                                       │
│   MCP server(s)                               │
│       ├─ Tools (actions)                      │
│       ├─ Prompts (templates)                  │
│       └─ Resources (read-only data)           │
│           │                                   │
│           ▼                                   │
│       External services / data                │
│                                               │
└───────────────────────────────────────────────┘

3 primitives

Decision framework

Build MCP server when:

• > 5 integrations planned
• Share tools across multiple apps
• Want community to use (open-source)
• Long-term maintenance
• Service provider exposing for AI use

Build MCP client when:

• Connect to existing MCP servers
• Enterprise integrating multiple systems
• Cross-app tool access needed

Skip MCP (use plain tool use):

• Simple app, 1-3 custom tools
• Proof of concept, MVP
• Tools deeply tied to app logic
• Short-term project

Skip API direct (use MCP):

• Slack/GitHub/etc. provide MCP → use MCP (less maintenance)
• You expose service to AI clients → MCP standard

Production checklist

Server side

• Tools have descriptive names + rich descriptions
• Pydantic Field descriptions cho mỗi param
• Error handling với raise ValueError("message")
• Auth/permission checks if sensitive
• Rate limiting if external API backend
• Logging qua log_level
• Tested qua Inspector + pytest
• Resource URIs follow consistent scheme
• Prompts well-crafted (expert-level)

Client side

• Reuse sessions (not per-request)
• Handle isError from tool calls
• Timeout on tool calls
• Tool name namespacing (if multi-server)
• Error recovery / retry
• UX for slash commands (prompts)
• Resource caching
• Monitor: tool call volume, success rate, latency

Integration with Claude

• MCP tools converted correctly to Anthropic schema
• Multi-turn loop with max_turns safety
• Tool results formatted correctly back to Claude
• Conversation history managed

Common mistakes

1. MCP server with side effects in resources

Violates contract. Debug hell.

Fix: Resources = read only. Side effects = tools.

2. Tool name collision between servers

Multiple servers → same tool name → ambiguity.

Fix: Namespace github:get_user, slack:get_user.

3. Not handling disconnection

Server crashes → client hangs.

Fix: Heartbeat, auto-reconnect.

4. Exposing too much

All DB tables as resources → security risk, confusion.

Fix: Curated views only.

5. Per-tool manual schema

Manually maintain JSON schemas → drift from code.

Fix: Use @mcp.tool + type hints.

Ecosystem resources

Official servers (curated list)

github.com/modelcontextprotocol/servers

Includes:

• Filesystem
• GitHub
• PostgreSQL
• Brave Search
• Google Drive
• Slack
• Memory
• Sequential thinking
• Many more

Community

• Python SDK: pip install mcp
• TypeScript SDK: npm install @modelcontextprotocol/sdk
• Docs: modelcontextprotocol.io
• Spec: github.com/modelcontextprotocol/specification

Claude Desktop

Native MCP integration. Install MCP servers, use directly trong Desktop chat.

Config: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/docs"]
    }
  }
}

Advanced topics (not covered in course)

• Authentication/OAuth — MCP 2.0+ spec
• Streaming responses — large data streams
• Logging / observability — structured MCP logs
• Compression — cho large resource content
• Federated MCP — one server aggregating others
• Security model — sandboxing untrusted servers

Anthropic docs có chi tiết.

When MCP không fit

• Latency critical (< 50ms) — MCP adds overhead
• Simple 1-tool integration — plain tool use simpler
• Ephemeral tools — tạo cho 1 conversation rồi bỏ

Review code: Full MCP chatbot

Bạn đã build:

• mcp_server.py — FastMCP với tools, prompts, resources
• mcp_client.py — Python client wrapper
• main.py — chatbot với Claude + MCP integration
• Inspector testing workflow
• Slash commands
• @mention resources

Production-ready skeleton. Real apps add:

• Auth
• Persistent storage
• UI (web, desktop)
• Deployment (containerize MCP server)

Áp dụng ngay

Bài tập 1: Evaluate your project (20 phút)

Review your MCP implementation:

• Production checklist pass?
• Any anti-patterns present?
• Performance acceptable?

Bài tập 2: Explore ecosystem (30 phút)

Visit github.com/modelcontextprotocol/servers. Try 1 official server (filesystem, sqlite, memory).

Connect qua your MCP client. Test.

Tóm tắt bài học